[igt-dev] [PATCH i-g-t v2 08/10] tests/i915/exec_fence: add timeline fence tests

[Lionel Landwerlin]

On 22/11/2019 18:11, Chris Wilson wrote:
> Quoting Lionel Landwerlin (2019-11-22 16:03:16)
>> On 22/11/2019 17:51, Chris Wilson wrote:
>>> Quoting Lionel Landwerlin (2019-11-22 15:43:11)
>>>> On 22/11/2019 17:26, Chris Wilson wrote:
>>>>> Quoting Lionel Landwerlin (2019-11-22 13:02:58)
>>>>>> We can now give a tuple (handle, point_value) for timeline semaphores.
>>>>>>
>>>>>> v2: Fix up syncobj-timeline-repeat test (Lionel)
>>>>>>        Update following kernel rename
>>>>>>
>>>>>> v3: Add more ordering tests for syncobj/buffer handle state (Chris)
>>>>> Another question, do we worry about sw_sync imported fences being
>>>>> handles differently to native?
>>>>> -Chris
>>>> What do you mean by differently?
>>> We have is_i915_fence() with branching paths for internal/external, and
>>> within the syncobj timeline mixed fence contexts cause a hard barrier,
>>> so may present interesting effects at the uAPI layer?
>>
>> Well... same issues with an array of i915 fences right?
>>
>>
>> I guess fence_set_priority() could be adjusted to deal with
>> dma-fence-chain like it does with dma-fence-array.
> No, either it's a chain of fence or it's just a list (and shouldn't be
> calling itself a chain ;). If it is a chain with dependencies along the
> chain, then we already have the information for priority inheritance.
> Except dma-fence-chain doesn't _guarantee_ that is the case and leaves
> it up to the caller to provide order across fence_contexts (within a
> fence_context it the seqno are meant to be ordered, and if the caller
> tries to use them out-of-order for syncpts within a dma-fence-chain it
> goes into the broken chain path same as for switching fence_contexts).
> -Chris

I could walk the chain and peel the dma-chain-fence nodes to add the 
actual underlying fences in the execbuf path.

But it sounds more like optimization than something else.


I'm not sure I follow your point about seqno ordering though.

Within a timeline it's always ordered (because of the new = MAX(old, new)).

Even if the context changes, the seqnos are always increasing.


-Lionel